Abstract
BACKGROUND: The left atrial appendage (LAA) is a critical but frequently overlooked site of thrombus formation, reinforcing the need for accurate identification in routine cardiac imaging. This process is related to pathological dilation associated with endothelial injury and a proinflammatory status. This study assesses the performance of deep learning architectures based on U-Net, specifically UNet3D, Residual-UNet3D, 3D Attention-UNet, and Res16-PAC-UNet, in the semiautomated segmentation and volume measurement of LAA. METHODS: We retrospectively analyzed noncontrast cardiac computed tomography (NCCT) scans from 452 patients aged ≥ 60 years, acquired for chest pain evaluation, to compare the performance of four U-Net-based deep learning architectures (UNet3D, Residual-UNet3D, 3D Attention-UNet, and Res16-PAC-UNet) for semiautomated LAA segmentation and volume measurement. Segmentation accuracy was assessed with the Dice coefficient, and volumetric agreement with Pearson correlation and Bland-Altman analysis. RESULTS: Dice coefficients were 78.44 ± 1.93 for UNet3D, 78.97 ± 0.79 for Residual-UNet3D, 79.07 ± 1.43 for 3D Attention-UNet, and 77.68 ± 1.47 for Res16-PAC-UNet. All models showed strong correlations between predicted and manual volumes (P < 0.001), with the highest in 3D Attention-UNet (r = 0.800). Bland-Altman analysis indicated minimal bias and narrow limits of agreement for all architectures, confirming consistent reliability. CONCLUSIONS: Deep learning-based segmentation on NCCT enables accurate, reproducible LAA morphological and volumetric assessment without contrast, offering a rapid and reliable tool to support cardiovascular risk stratification and treatment planning.